fix(backend): use threading.local() for per-thread Redis mutex

Replace event-loop detection in locks() with threading.local() storage.
The previous approach had a TOCTOU race: two threads could interleave
the check-and-set of _locks/_locks_loop, causing one thread to use a
mutex created on another thread's event loop.

threading.local() gives each thread its own mutex naturally — no races,
no shared mutable state, no stale-loop checks needed.